Merge task branch for: fix(worker): planning-chain cascade stalls at an Idle middle link

src/ClaudeDo.Worker/Planning/PlanningChainCoordinator.cs

@@ -92,14 +92,7 @@ public sealed class PlanningChainCoordinator
     public async Task<string?> OnChildFinishedAsync(
         string childTaskId, TaskStatus finalStatus, CancellationToken ct = default)
     {
-        await using var ctx = await _dbFactory.CreateDbContextAsync(ct);
-        // The successor is whichever sibling explicitly blocks on this child.
-        var nextId = await ctx.Tasks
-            .AsNoTracking()
-            .Where(t => t.BlockedByTaskId == childTaskId)
-            .OrderBy(t => t.SortOrder).ThenBy(t => t.CreatedAt)
-            .Select(t => t.Id)
-            .FirstOrDefaultAsync(ct);
+        var nextId = await FindSuccessorAsync(childTaskId, ct);
         if (nextId is null) return null;
 
         if (finalStatus == TaskStatus.Done)
@@ -108,10 +101,33 @@ public sealed class PlanningChainCoordinator
             return nextId;
         }
 
-        // Child failed or was cancelled: cancel the immediate successor so the chain
-        // is not left wedged. CancelAsync triggers OnChildTerminalAsync → OnChildFinishedAsync
-        // for that successor, cascading cancellation through the rest of the chain.
-        await _state().CancelAsync(nextId, DateTime.UtcNow, ct);
-        return null;
+        // Child failed or was cancelled: cancel the immediate successor so the chain is
+        // not left wedged. If it's cancellable, CancelAsync's own OnChildTerminalAsync
+        // callback recurses into this method for its successor, cascading cancellation
+        // through the rest of the chain. If it's not (e.g. it was parked back to Idle
+        // out of band), CancelAsync is a no-op and nothing will ever call back for it —
+        // keep walking the chain ourselves so the tail isn't left wedged forever.
+        var predecessorId = nextId;
+        while (true)
+        {
+            var result = await _state().CancelAsync(predecessorId, DateTime.UtcNow, ct);
+            if (result.Ok) return null;
+
+            var following = await FindSuccessorAsync(predecessorId, ct);
+            if (following is null) return null;
+            predecessorId = following;
+        }
+    }
+
+    // The successor is whichever sibling explicitly blocks on this task.
+    private async Task<string?> FindSuccessorAsync(string taskId, CancellationToken ct)
+    {
+        await using var ctx = await _dbFactory.CreateDbContextAsync(ct);
+        return await ctx.Tasks
+            .AsNoTracking()
+            .Where(t => t.BlockedByTaskId == taskId)
+            .OrderBy(t => t.SortOrder).ThenBy(t => t.CreatedAt)
+            .Select(t => t.Id)
+            .FirstOrDefaultAsync(ct);
     }
 }

tests/ClaudeDo.Worker.Tests/Planning/PlanningChainCoordinatorTests.cs

@@ -178,6 +178,37 @@ public sealed class PlanningChainCoordinatorTests : IDisposable
         Assert.Equal(TaskStatus.Cancelled, kids[3].Status);
     }
 
+    [Fact]
+    public async Task OnChildFailed_MidChain_IdleSuccessor_DoesNotWedgeTail()
+    {
+        // Chain: c0 → c1 → c2 → c3. c1 is parked back to Idle (e.g. a manual reset)
+        // while still wired into the chain (BlockedByTaskId = c0). c0 then fails:
+        // CancelAsync(c1) is a no-op because Idle isn't a cancellable state, so the
+        // cascade must keep walking past c1 instead of leaving c2/c3 wedged forever.
+        await SeedPlanningFamilyAsync("P", 4);
+        await _sut.SetupChainAsync("P", enqueue: true, default);
+
+        await using (var ctx = _factory.CreateDbContext())
+        {
+            var c0 = await ctx.Tasks.FirstAsync(t => t.Id == "P-c0");
+            c0.Status = TaskStatus.Failed;
+            var c1 = await ctx.Tasks.FirstAsync(t => t.Id == "P-c1");
+            c1.Status = TaskStatus.Idle;
+            await ctx.SaveChangesAsync();
+        }
+
+        var advanced = await _sut.OnChildFinishedAsync("P-c0", TaskStatus.Failed, default);
+
+        Assert.Null(advanced);
+        var kids = await GetChildrenAsync("P");
+        Assert.Equal(TaskStatus.Failed, kids[0].Status);
+        // c1 was already parked Idle; the coordinator must not resurrect it.
+        Assert.Equal(TaskStatus.Idle, kids[1].Status);
+        // c2/c3 must not stay wedged as Queued+blocked forever.
+        Assert.Equal(TaskStatus.Cancelled, kids[2].Status);
+        Assert.Equal(TaskStatus.Cancelled, kids[3].Status);
+    }
+
     [Fact]
     public async Task OnChildDone_LastChild_ReturnsNull()
     {